intra: 3d intracranial aneurysm dataset for deep...
TRANSCRIPT
IntrA: 3D Intracranial Aneurysm Dataset for Deep Learning
Xi Yang1 Ding Xia1,2 Taichi Kin1 Takeo Igarashi1
1The University of Tokyo 2South China University of Technology
Figure 1: 3D models of intracranial aneurysm segments with segmentation annotation in our dataset. Hot pink shows the
healthy blood vessel part, and aqua shows the aneurysm part for each model.
Abstract
Medicine is an important application area for deep
learning models. Research in this field is a combina-
tion of medical expertise and data science knowledge. In
this paper, instead of 2D medical images, we introduce
an open-access 3D intracranial aneurysm dataset, IntrA,
that makes the application of points-based and mesh-based
classification and segmentation models available. Our
dataset can be used to diagnose intracranial aneurysms and
to extract the neck for a clipping operation in medicine
and other areas of deep learning, such as normal estima-
tion and surface reconstruction. We provide a large-scale
benchmark of classification and part segmentation by test-
ing state-of-the-art networks. We also discuss the perfor-
mance of each method and demonstrate the challenges of
our dataset. The published dataset can be accessed here:
https://github.com/intra3d2019/IntrA.
1. Introduction
Intracranial aneurysm is a life-threatening disease, and
its surgical treatments are complicated. Timely diagnosis
and preoperative examination are necessary to formulate
the treatment strategies and surgical approaches. Currently,
the primary treatment method is clipping the neck of an
aneurysm to prevent it from rupturing, as shown in Figure 2.
The decisions of the position and posture of the clip are still
highly dependent on “clinical judgment” based on the ex-
perience of physicians. In the surgery support system of in-
tracranial aneurysms simulating real-life neurosurgery and
teach neurosurgical residents [5], the accuracy of aneurysm
segmentation is the most crucial part because it is used to
extract the neck of an aneurysm, that is, the boundary line
of the aneurysm.
Based on 3D surface models, the diagnosis of an
aneurysm can be much more accurate than 2D images. The
edge of the aneurysm is much clearer for doctors, and the
complicated and time-consuming annotation of a mess of
2D images is avoided. There are many reports of automatic
diagnosis and segmentation of aneurysms based on medical
images, including intracranial aneurysm (IA) and abdomi-
nal aortic aneurysm (AAA) [24, 30, 37]; however, few re-
ports have been published based on 3D models. This is not
only because data collection is inefficient, subjective, and
2656
Figure 2: The treatment of intracranial aneurysm by clip-
ping.
challenging to share in medicine, but also the joint knowl-
edge of computer application science and medical science.
Objects with arbitrary shapes are ubiquitous, and a non-
Euclidean manifold reveals more critical information than
using Euclidean geometry, like complex typologies of brain
tissues in neuroscience[7]. However, the study of 2D mag-
netic resonance angiography (MRA) images confines the
selection to 3D neural networks based on pixels and voxels,
which also omits the information from manifolds. There-
fore, we propose an open-access 3D intracranial aneurysm
dataset to solve the above issues and to promote the applica-
tion of deep learning models in medical science. The points
and meshes-based models exhibit excellent generalization
abilities for 3D deep learning tasks in our experiments.
Our main contributions are:
1. We propose an open dataset that consists of 3D
aneurysm segments with segmentation annotations,
automatically generated blood vessel segments, and
complete models of scanned blood vessels of the
brains. All annotated aneurysm segments are pro-
cessed as manifold meshes.
2. We develop tools to generate 3D blood vessel segments
from complete models and to annotate a 3D aneurysm
model interactively. The data processing pipeline is
also introduced.
3. We evaluate the performance of various state-of-the-
art 3D deep learning methods on our dataset to pro-
vide benchmarks of classification (diagnose) and seg-
mentation of intracranial aneurysms. Furthermore, we
analyze the different features of each method from the
results obtained.
2. Related Work
2.1. Datasets
Medical dataset. Large-scale samples are required to
surmount the challenges of the complexity and heterogene-
ity of many diseases, but data collection in medical research
is costly, it is unattainable for a single research institute.
Therefore, data sharing is critical. Several medical datasets
have been published online for collaboration on finding a
treatment solution. For example, integrated dataset Med-
Pix [4], bone X-rays dataset MURA [35], brain neuroimag-
ing dataset [15], Medical Segmentation Decathlon [38],
Harvard GSP [8], and SCR database [42]. Data collec-
tion is also critical for a single category of disease, such as,
The Lung Image Database Consortium (LIDC-IDRI) [3],
Indian Diabetic Retinopathy Image Dataset (IDRiD) [2],
EyePACS [2], and Autism Brain Imaging Data Exchange
(ABIDE) [1]. To date, almost all of them are 2D medical
images.
Non-medical 3D dataset. In recent years, 3D model
datasets were introduced in the research of computer vi-
sion and computer graphics with the development of deep
learning algorithms. For instance, CAD model datasets:
modelNet [48], shapeNet [9], COSEG Dataset [44], ABC
dataset [23]; 3D printing model datasets: Thingi10K [52],
Human model dataset [31], etc. Various 3D deep learning
tasks are widely carried out on these datasets.
2.2. Methods
A 3D model has four kinds of representations, projected
view, voxel, point cloud, and mesh. The methods based
on projected view or voxel are implemented conveniently
using similar structures with 2D convolutional neural net-
works (CNNs). Point cloud or mesh has a more accurate
representation of a 3D shape; however, new convolution
structures are required.
Projected View. Su et al. proposed a multi-view CNN
to recognize 3D shapes [40]. Kalogerakis et al. com-
bined image-based fully convolutional networks (FCNs)
and surface-based conditional random fields (CRFs) to yield
coherent segmentation of 3D shapes [21].
Voxel. Cicek et al. introduced 3D U-Net for volumetric
segmentation that learns from sparsely annotated volumet-
ric images [10]. Wang et al. presented O-CNN, an Octree-
based Convolutional Neural Network (CNN), for 3D shape
analysis [43]. Graham et al. designed new sparse convolu-
tional operations to process spatially-sparse 3D data, called
submanifold sparse convolutional networks (SSCNs) [16].
Wang and Lu proposed VoxSegNet to extract discrimina-
tive features encoding detailed information under limited
resolution [46]. Le and Duan proposed the PointGrid, a
3D convolutional network that is an integration of point and
grid [26].
Points. Qi et al. proposed PointNet, making it is possible
to input 3D points directly for neural work [33], then they
introduced a hierarchical network PointNet++ to learn local
features [34]. Based on these pioneering works, many new
convolution operations were proposed. Wu et al. treated
convolution kernels as nonlinear functions of the local coor-
2657
Complete model (103) Generated segments (1909) Annotated segments (116)
Figure 3: There are three types of data in our dataset. The automatically generated blood vessel segments can be with or
without aneurysms, and an aneurysm is often partially separated. We see that both aneurysm segments and healthy vessel
segments have complex shapes with a different number of branches. For example, in the annotated segments, the aneurysm
is huge in the second segment of the first row, but it is tiny in the first segment of the third row. The second segment of the
third row has even two aneurysms.
dinates of 3D points comprised of weight and density func-
tions, named PointConv [47]. Li et al. presented PointCNN
that can leverage spatially local correlation in data repre-
sented densely in grids for feature learning [28]. Xu et al.
designed the filter as a product of a simple step function that
captures local geodesic information and a Taylor polyno-
mial, named SpiderCNN [49]. Moreover, the SO-Net mod-
els the spatial distribution of point cloud by building a Self-
Organizing Map (SOM) [27]. Su et al. presented SPLAT-
Net for processing point clouds that directly operates on a
collection of points represented as a sparse set of samples
in a high-dimensional lattice [39]. Zhao et al. proposed 3D
point-capsule networks [51]. Wang et al. proposed dynamic
graph neural network (DGCNN) [45]. Yang et al. devel-
oped Point Attention Transformers (PATs) to process the
point clouds [50]. Thomas et al. presented a new design of
point convolution, called Kernel Point Convolution1 (KP-
Conv) [41]. Liu et al. proposed a dynamic points agglomer-
ation module to construct an efficient hierarchical point sets
learning architecture. [29].
Mesh. Maron et al. applied a convolution operator to
sphere-type shapes using a global seamless parameteriza-
tion to a planar flat-torus [31]. Hanocka et al. utilize the
unique properties of the triangular mesh for direct analysis
of 3D shapes, named MeshCNN [17]. Feng et al. regard the
polygon faces as the unit, split their features into spatial and
structural features called MeshNet [14].
3. Our Dataset
3.1. Data
Our dataset includes complete models with aneurysms,
generated vessel segments, and annotated aneurysm seg-
ments, as shown in Figure 3. 103 3D models of entire
brain vessels are collected by reconstructing scanned 2D
MRA images of patients. We do not publish the raw 2D
MRA images because of medical ethics. 1909 blood ves-
sel segments are generated automatically from the complete
models, including 1694 healthy vessel segments and 215
aneurysm segments for diagnosis. An aneurysm can be di-
vided into segments that can verify the automatic diagnosis.
116 aneurysm segments are divided and annotated manu-
ally by medical experts; the scale of each aneurysm seg-
ment is based on the need for a preoperative examination.
The details are described in the next section. Furthermore,
geodesic distance matrices are computed and included for
each annotated 3D segment, because the expression of the
geodesic distance is more accurate than Euclidean distance
according to the shape of vessels. The matrix is saved as
N × N for a model with N points, shortening the training
computation time.
Our data have several characteristics common to med-
ical data: 1) Small but diverse. The amount of data
is not so large compared to other released CAD model
datasets; however, it includes diverse shapes and scales of
intracranial aneurysms as well as different amounts of ves-
2658
Figure 4: An intracranial aneurysm is restored from incom-
plete scanned data by neurosurgeons.
sel branches. 2) Unbalanced. The number of points of
aneurysms and healthy vessel parts is imbalanced based on
the shape of aneurysms. The number of 3D aneurysm seg-
ments and healthy vessel segments are not equal because the
aneurysms are usually much smaller than the entire brain.
Challenge. Experts collected out dataset instead of regu-
lar people. Intact 3D models have to be restored from recon-
structed data manually, as shown in Figure 4. Besides, the
annotation of the neck of aneurysms necks requires years
of clinical experience in complex situations. Also, the 3D
models are not manifold. We clean the surface meshes to
create an ideal dataset for algorithm research.
Statistics and analysis. The statistics of our dataset
are shown in Figure 5. We count the number points in
each segment to some extent express the difference of the
shapes, since the points are mostly uniformly distributed on
the surface. The number of points generated in 1909 seg-
ments is approximately 500 to 1700 at Geodesic distance
30. Our dataset includes all types of intracranial aneurysms
in medicine: bifurcation type, trunk type, blister type, and
combined type. The shapes of aneurysm are diverse in
our dataset both in geometry and topology; six aneurysm
segments selected are shown in the right of Figure 3. Be-
sides, we calculated the size of an aneurysm as the ratio be-
tween the diagonal distance of the global segment and the
aneurysm part instead of the real size of the parent vessel or
the aneurysm.
3.2. Tools
We developed annotation tools and segment generation
tools to assist in constructing our dataset.
Annotation. Users draw an intended boundary by click-
ing several points. The connection between two points is
determined by the shortest path. After users create a closed
boundary line, they annotate the aneurysm part by select-
ing a point inside of it. The enclosed area is calculated au-
tomatically by propagation from the point to the boundary
line along with surface meshes. With the support of multi-
ple boundary lines, the annotation tool also can be used for
separating both the aneurysm part and the aneurysm seg-
ment manually, as shown in Figure 6.
Vessel segment generation. Vessel segments are gen-
erated by randomly picking points from the complete mod-
els and selecting the neighbor area whose geodesic distance
along the vessel is smaller than a threshold. We also man-
ually select points for increasing the number of segments
with an aneurysm. To construct an ideal dataset, few data
which are ambiguous or only include trivial components are
removed by using our visualization tool.
3.3. Processing pipeline
The processing pipeline is shown in Figure 7, more de-
tails are described in the supplementary material.
3D reconstruction and restore. Our data are ac-
quired by Time-Of-Flight Magnetic Resonance Angiogra-
phy (TOF-MRA) of human brain. Using the single thresh-
old method [19], each complete 3D model is reconstructed
from 512 × 512 × 180 ∼ 300 2D images sliced by
0.469 × 0.469 × 1mm. The aneurysm segments are sep-
arated and restored interactively using the multi-threshold
method [22] by two neurosurgeons, then processed by
Gaussian smoothing. This image processing is conducted
in life sciences software, Amira 2019 (Thermo Fisher Sci-
entific, MA, USA). It takes about 50 workdays in total.
Generation and annotation. By using our generation
and annotation tools, blood vessel segments are obtained
and classified. The segmentation annotation of aneurysm
segments is also finished. A neurosurgeon completed it in 8
hours.
Data clean and re-meshing. The reconstructed 3D
models are noisy and not manifold. Huang et al. [20] de-
scribed an algorithm to generate a manifold surface for
3D models; however, this method does not remove iso-
lated components and significantly changes the shape of the
model. Therefore, we use filter tools in MeshLab to remove
duplicate faces and vertices, and separate pieces in the data
manually, which ensure that the models do not have non-
manifold edges. MeshLab also generates the normal vector
at each point. The geodesic matrix is computed by solving
the heat equation on the surface using a fast approximate
geodesic distance method by [11].
3.4. Supported Studies
Diagnose (Classification). The diagnosis of an
aneurysm can be considered as a classification problem of
aneurysms and healthy vessel segments. From a 3D brain
model of a patient, vessel segments are generated by our
tools; then, the diagnosis is completed by classifying the
segments with aneurysms.
Part segmentation. Our annotated 3D models present a
precise boundary of each aneurysm to support segmentation
research. The data is easy to convert to any 3D representa-
tion of various deep learning algorithms.
Rule-based algorithms. Besides, algorithms based on
rules are proposed for either aneurysm in the brain or ab-
2659
Figure 5: Statistics of annotated models.
Figure 6: The top figure shows the UI of our annotation tool.
The points the user clicked are shown as blue to decide the
boundary of the aneurysm, then a random point (yellow)
is selected to annotate the aneurysm part (green). The two
bottom figures demonstrate the results of multiple boundary
lines.
dominal aortic aneurysm [13, 25, 36]. The accuracy and
generalization of the designed rules should be verified by
large data.
Others. Because we provide the information about the
normal vector and geodesic distance in our dataset, other
researches on 3D models are also supported, e.g. normal
estimation [6], geodesic estimation [18], 3D surface recon-
struction [12, 32], and etc.
4. Benchmark
We selected state-of-the-art methods as the benchmarks
of classification and segmentation of our dataset. We im-
plemented dataset interfaces to the original implementa-
tions by the authors and kept the same hyper-parameters
and loss functions of models as in the original papers. A
detailed explanation of the implementation of each method
is described in the supplementary material. We tested
these methods by 5-fold cross-validation. The shuffled data
was divided into 5 subsamples and was the same for each
method. 4 subsamples were used as training data, 1 was for
test data. The experiments were carried out on PCs with
GeForce RTX 2080 Ti ×2, GeForce GTX 1080 Ti ×1. The
net training time of all methods was over 92 hours.
4.1. Classification
6 methods were selected for binary classification bench-
marks, including PointNet [33], PointNet++ (PN++) [34],
PointCNN [28], SpiderCNN [49], self-organizing network
(SO-Net) [27], dynamic graph CNN (DGCNN) [45]. We
combined the generated blood vessel segments and man-
ually segmented aneurysms in total 2025 as the dataset
for testing classification accuracy and F1-Score of each
method. The experimental results are shown in Table 1.
PN++ with 1024 sampling points has the highest ac-
curacy of aneurysms, and PointCNN with 2048 sampled
points has the greatest accuracy of artery and F1-Score.
The accuracy and F1-score of almost all methods showed
an increasing tendency as more sampled points were pro-
vided. However, SpiderCNN attained the highest aneurysm
detection rate and F1-Score at 1024 sampling points. The
majority of misclassified 3D models contained small-sized
or incomplete aneurysms that are hard to distinguish from
healthy blood vessels.
4.2. Segmentation
We selected 11 networks, PointGrid [26], two kind of
submanifold sparse convolutional networks (SSCNs): fully-
connected (SSCN-F) and Unet-like (SSCN-U) [16] struc-
tures, PointNet [33], two kind of PointNet++ [34]: input
with normal (PN++) and with normal and geodesic dis-
tance (PN++g), PointConv [47], PointCNN [28], Spider-
CNN [49], MeshCNN [17], and SO-Net [27], to provide
segmentation benchmarks. 116 annotated aneurysm seg-
ments were used for evaluating these methods. The test of
2660
Figure 7: Pipeline of data processing. We generate 3D surface models of brain blood vessels from MRA images. Then, the
segments are generated and annotated. Data clean and re-meshing are conducted as the requirement of mesh-based methods.
Network Input V. (%). A. (%) F1-Score
PN++
512 98.52 86.69 0.8928
1024 98.52 88.51 0.9029
2048 98.76 87.31 0.9016
SpiderCNN
512 98.05 84.58 0.8692
1024 97.28 87.90 0.8722
2048 97.82 84.89 0.8662
SO-Net
512 98.76 84.24 0.8840
1024 98.88 81.21 0.8684
2048 98.88 83.94 0.8850
PointCNN
512 98.38 78.25 0.8494
1024 98.79 81.28 0.8748
2048 98.95 85.81 0.9044
DGCNN
512/10 95.22 60.73 0.6578
1024/20 95.34 72.21 0.7376
2048/40 97.93 83.40 0.8594
PointNet
512 94.45 67.66 0.6909
1024 94.98 64.96 0.6835
2048 93.74 69.50 0.6916
Table 1: Classification results of each method. The second
column shows the number of input points, the additional
input K is required for DGCNN. The accuracies of healthy
vessel segments (V.) and aneurysm segments (A.) and F1-
Score are calculated by the mean value of each fold.
each subsample was repeated 3 times, and the final results
were the mean values of each best result. We assessed the
effects of each method using two indexes: Jaccard Index
(JI) and Sørensen-Dice Coefficient (DSC). Jaccard Index is
also known as Intersection over Union (IoU). The results of
segmentation are shown in Figure 8 and Table 2.
Methods based on points. The segmentation meth-
ods based on point cloud obtained good results and main-
tained the same level with the results on ShapeNet [9]. SO-
Net showed excellent performance on IOU and DSC of
aneurysms, while PointConv had the best result on parent
blood vessels. PN++ had the third-best performance and
had the fastest training speed (5s per epoch, and converged
at approximately an epoch of 115 on GTX 1080 Ti). Mean-
while, PointCNN had the slowest training speed (24s per
epoch, and converged at approximately an epoch of 500 on
GTX 1080 Ti) and moderate segmentation accuracy. Spi-
derCNN did not have the same performance as it had on
the ShapeNet, but CI95 was unusually high. Besides the
methods mentioned in Section 4.2, we also tried 3D Cap-
suleNet [51], but it classified every point into the healthy
blood vessel, which shows its limited generalization cross-
ing datasets.
Resolution of voxels. Methods based on voxels
achieved relatively low IOU and DSC on each fold. The
performance of SSCN grew as the resolution was increased
from 24 to 40 (the resolution 24 was offered by the authur
in the code). But the average IOU had a fluctuation of about
8%, which was quite obvious compared to other methods
(about 2%). Based on the paper of PointGrid, N = 16 and
K = 2 were recommended parameters. However, we no-
ticed that the combination of N = 32 and K = 2 achieved
the highest scores.
Common poorly segmented 3D models. Most mod-
els were segmented excellent as top two rows in Figure 9.
However, the accuracy dropped when the aneurysm occu-
pied a small size ratio of the segment, like the third and
fourth row. Meanwhile, the segmentation performance of
aneurysms with a large size ratio was satisfactory. The fifth
row shows a special segment with 2 aneurysms. Although
most of methods failed to segment it, PointConv and PN++
with geodesic information maintained a good performance.
Geodesic information. Compared to other CAD model
datasets, the complex shapes of blood vessels is a different
challenge in part segmentation. Methods based on points
usually use the Euclidean distance to estimate the relevance
between points. However, it is not ideal for our dataset. For
example, PN++ misclassified the aneurysm points close to
the blood vessels even with normal information, as shown in
the last row of Figure 9. While, by using geodesic distance,
2661
Figure 8: IoU results of the aneurysm part for each fold of networks. These methods are compared with their best perfor-
mance.
Goundtruth PointConv PN++g PN++
Figure 9: Results comparison of three networks. More de-
tails are described in the supplementary material.
PN++ learned more exact spatial structure. PointConv also
segmented it well. Its excellent performance can be at-
tributed to the network learning the parameters of spatial
filters. In addition, MeshCNN segmented every aneurysm
decently although the overall performance is not best, which
owes to its convolution on meshes providing information on
manifolds.
5. Conclusion and Further Work
In this paper, we introduced a 3D dataset of intracranial
aneurysm with annotation by experts for geometric deep
learning networks. The developed tools and data process-
ing pipeline is also released. Furthermore, we evaluated and
analyzed the state-of-the-art methods of 3D object classifi-
cation and part segmentation on our dataset. The existing
methods are likely to be less effective on complex objects,
though they perform well on the segmentation of common
ones. It is possible to improve further the performance and
generalization of networks when geodesic or connectivity
information on 3D surfaces is accessible. The introduction
of our dataset can be instructive to the development of new
structures of geometric deep learning methods for medical
datasets.
In further work, we will keep increasing processed real
data for our dataset. Besides, we will verify the feasibility
of synthetic data for data augmentation, which can signifi-
cantly improve the efficiency of data collection. We hope
more deep learning networks will be applied to medical
practice.
Acknowledgements
This research was supported by AMED under Grant
Number JP18he1602001.
2662
Network InputIoU (%) CI 95 (%) DSC (%) CI 95 (%)
V. A. V. A. V. A. V. A.
Po
int
SO-Net
512 94.22 80.14 92.12 ∼ 96.32 73.71 ∼ 86.57 96.95 87.90 95.79 ∼ 98.12 83.14 ∼ 92.66
1024 94.42 80.99 92.39 ∼ 96.45 74.71 ∼ 87.27 97.06 88.41 95.93 ∼ 98.19 83.65 ∼ 93.17
2048 94.46 81.40 92.51 ∼ 96.41 75.37 ∼ 87.43 97.09 88.76 96.02 ∼ 98.16 84.38 ∼ 93.15
Po
int
PointConv
512 94.16 79.09 91.76 ∼ 96.56 70.26 ∼ 87.92 96.89 86.01 95.55 ∼ 98.24 78.34 ∼ 93.69
1024 94.59 79.42 92.53 ∼ 96.66 70.55 ∼ 88.29 97.15 86.29 96.00 ∼ 98.30 78.33 ∼ 94.25
2048 94.65 79.53 92.64 ∼ 96.67 70.96 ∼ 88.10 97.18 86.52 96.06 ∼ 98.30 78.95 ∼ 94.09
Po
int
PN++g
512 93.34 75.74 91.07 ∼ 95.60 66.51 ∼ 84.97 96.47 83.90 95.19 ∼ 97.74 75.68 ∼ 92.12
1024 93.28 76.53 90.93 ∼ 95.62 67.91 ∼ 85.15 96.43 84.82 95.12 ∼ 97.75 77.65 ∼ 91.99
2048 93.60 76.95 91.44 ∼ 95.75 68.76 ∼ 85.14 96.62 85.18 95.41 ∼ 97.82 78.39 ∼ 91.98
Po
int
PN++
512 93.42 76.22 90.91 ∼ 95.92 66.70 ∼ 85.73 96.48 83.92 95.04 ∼ 97.92 75.46 ∼ 92.38
1024 93.35 76.38 91.10 ∼ 95.60 67.96 ∼ 84.80 96.47 84.62 95.20 ∼ 97.74 77.45 ∼ 91.78
2048 93.24 76.21 90.93 ∼ 95.56 67.99 ∼ 84.43 96.40 84.64 95.08 ∼ 97.72 77.71 ∼ 91.57
Po
int
PointCNN
512 92.49 70.65 89.77 ∼ 95.22 58.89 ∼ 82.42 95.97 78.55 94.41 ∼ 97.54 67.37 ∼ 89.73
1024 93.47 74.11 91.11 ∼ 95.84 63.54 ∼ 84.68 96.53 81.74 95.20 ∼ 97.86 71.88 ∼ 91.59
2048 93.59 73.58 91.45 ∼ 95.73 62.81 ∼ 84.35 96.62 81.36 95.43 ∼ 97.80 71.39 ∼ 91.33
Mes
h
MeshCNN
750 85.43 55.63 81.22 ∼ 89.64 46.53 ∼ 64.73 91.71 68.65 88.86 ∼ 94.56 60.16 ∼ 77.14
1500 90.86 71.32 88.20 ∼ 93.53 65.01 ∼ 77.62 95.10 82.21 93.55 ∼ 96.64 77.26 ∼ 87.16
2250 90.34 71.60 87.34 ∼ 93.34 63.99 ∼ 79.21 94.77 81.87 93.01 ∼ 96.53 75.72 ∼ 88.01
Po
int
SpiderCNN
512 90.16 67.25 86.34 ∼ 93.98 55.21 ∼ 79.29 94.53 75.82 92.17 ∼ 96.88 64.07 ∼ 87.56
1024 87.95 61.60 83.65 ∼ 92.24 48.97 ∼ 74.23 93.24 71.08 90.56 ∼ 95.93 58.50 ∼ 83.67
2048 87.02 58.32 82.57 ∼ 91.47 45.21 ∼ 71.44 92.71 67.74 89.94 ∼ 95.47 54.50 ∼ 80.98
Vo
xel
SSCN-F
24 87.95 56.56 84.32 ∼ 91.57 43.29 ∼ 69.84 93.35 66.04 91.14 ∼ 95.56 52.28 ∼ 79.80
32 88.08 55.26 84.62 ∼ 91.54 41.39 ∼ 69.13 93.44 64.05 91.36 ∼ 95.53 49.45 ∼ 78.66
40 90.09 61.45 87.00 ∼ 93.17 48.54 ∼ 74.37 94.62 70.54 92.78 ∼ 96.46 57.16 ∼ 83.91
Vo
xel
SSCN-U
24 87.43 55.78 83.48 ∼ 91.37 42.07 ∼ 69.48 93.00 65.03 90.58 ∼ 95.43 50.91 ∼ 79.16
32 86.13 53.52 82.12 ∼ 90.13 40.64 ∼ 66.40 92.24 64.01 89.72 ∼ 94.76 50.95 ∼ 77.06
40 88.66 57.94 85.25 ∼ 92.07 44.92 ∼ 70.96 93.78 67.39 91.73 ∼ 95.83 54.09 ∼ 80.64
Vo
xel
PointGrid
16/2 78.32 35.82 73.53 ∼ 83.12 25.22 ∼ 46.42 87.36 47.33 84.12 ∼ 90.60 35.28 ∼ 59.38
16/4 79.49 38.23 74.54 ∼ 84.44 26.29 ∼ 50.17 88.08 49.14 84.73 ∼ 91.43 35.65 ∼ 62.64
32/2 80.11 42.42 75.60 ∼ 84.62 30.51 ∼ 54.34 88.50 53.52 85.54 ∼ 91.45 40.41 ∼ 66.63
Po
int
PointNet
512 73.99 37.30 67.43 ∼ 80.56 26.17 ∼ 48.44 84.05 48.96 79.31 ∼ 88.79 36.53 ∼ 61.38
1024 75.23 37.07 69.10 ∼ 81.36 25.66 ∼ 48.48 85.00 48.38 80.69 ∼ 89.31 35.63 ∼ 61.13
2048 74.22 37.75 67.85 ∼ 80.60 26.85 ∼ 48.64 84.17 49.59 79.56 ∼ 88.78 37.48 ∼ 61.70
Table 2: Segmentation results of each network. The second column shows the number of input points, edges, or resolutions
for the methods based on points, mesh, and voxel, respectively. For PointGrid, the additional refers to the parameter K in
the paper. The healthy vessel part and aneurysm part are noted as V. and A., respectively. CI95 indicated 95% confidence
interval of IoU or DSC. The results are calculated by the mean value of each fold.
2663
References
[1] Autism Brain Imaging Data Exchange (ABIDE).
http://fcon_1000.projects.nitrc.org/
indi/abide/. 2
[2] Indian Diabetic Retinopathy Image Dataset (IDRiD).
https://idrid.grand-challenge.org/. 2
[3] LIDC-IDRI. https://wiki.
cancerimagingarchive.net/display/Public/
LIDC-IDRI. 2
[4] Medpix. https://medpix.nlm.nih.gov/home. 2
[5] Ali Alaraj, Cristian J Luciano, Daniel P Bailey, Abdussalam
Elsenousi, Ben Z Roitberg, Antonio Bernardo, P Pat Baner-
jee, and Fady T Charbel. Virtual reality cerebral aneurysm
clipping simulation with real-time haptic feedback. Opera-
tive Neurosurgery, 11(1):52–58, 2015. 1
[6] Alexandre Boulch and Renaud Marlet. Deep learning for
robust normal estimation in unstructured point clouds. In
Computer Graphics Forum, volume 35, pages 281–290. Wi-
ley Online Library, 2016. 5
[7] Michael M Bronstein, Joan Bruna, Yann LeCun, Arthur
Szlam, and Pierre Vandergheynst. Geometric deep learning:
going beyond euclidean data. IEEE Signal Processing Mag-
azine, 34(4):18–42, 2017. 2
[8] Randy L. Buckner, Joshua L. Roffman, and Jordan W.
Smoller. Brain Genomics Superstruct Project (GSP), 2014.
2
[9] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat
Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Mano-
lis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi,
and Fisher Yu. ShapeNet: An Information-Rich 3D Model
Repository. Technical Report arXiv:1512.03012 [cs.GR],
Stanford University — Princeton University — Toyota Tech-
nological Institute at Chicago, 2015. 2, 6
[10] Ozgun Cicek, Ahmed Abdulkadir, Soeren S Lienkamp,
Thomas Brox, and Olaf Ronneberger. 3d u-net: learning
dense volumetric segmentation from sparse annotation. In
International conference on medical image computing and
computer-assisted intervention, pages 424–432. Springer,
2016. 2
[11] Keenan Crane, Clarisse Weischedel, and Max Wardetzky.
Geodesics in heat: A new approach to computing distance
based on heat flow. ACM Transactions on Graphics (TOG),
32(5):152, 2013. 4
[12] Angela Dai, Charles Ruizhongtai Qi, and Matthias Nießner.
Shape completion using 3d-encoder-predictor cnns and
shape synthesis. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pages 5868–
5877, 2017. 5
[13] Sarada Prasad Dakua, Julien Abinahed, and Abdulla Al-
Ansari. A pca-based approach for brain aneurysm segmen-
tation. Multidimensional Systems and Signal Processing,
29(1):257–277, 2018. 5
[14] Yutong Feng, Yifan Feng, Haoxuan You, Xibin Zhao, and
Yue Gao. Meshnet: Mesh neural network for 3d shape rep-
resentation. In Proceedings of the AAAI Conference on Arti-
ficial Intelligence, volume 33, pages 8279–8286, 2019. 3
[15] Anthony F Fotenos, AZ Snyder, LE Girton, JC Morris, and
RL Buckner. Normative estimates of cross-sectional and lon-
gitudinal brain volume decline in aging and ad. Neurology,
64(6):1032–1039, 2005. 2
[16] Benjamin Graham, Martin Engelcke, and Laurens van der
Maaten. 3d semantic segmentation with submanifold sparse
convolutional networks. In Proceedings of the IEEE Con-
ference on Computer Vision and Pattern Recognition, pages
9224–9232, 2018. 2, 5
[17] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar
Fleishman, and Daniel Cohen-Or. Meshcnn: a network with
an edge. ACM Transactions on Graphics (TOG), 38(4):90,
2019. 3, 5
[18] Tong He, Haibin Huang, Li Yi, Yuqian Zhou, Chihao Wu,
Jue Wang, and Stefano Soatto. Geonet: Deep geodesic net-
works for point cloud analysis. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition,
pages 6888–6897, 2019. 5
[19] Romhild M Hoogeveen, Chris JG Bakker, and Max A
Viergever. Limits to the accuracy of vessel diameter mea-
surement in mr angiography. Journal of Magnetic Resonance
Imaging, 8(6):1228–1235, 1998. 4
[20] Jingwei Huang, Hao Su, and Leonidas Guibas. Robust water-
tight manifold surface generation method for shapenet mod-
els. arXiv preprint arXiv:1802.01698, 2018. 4
[21] Evangelos Kalogerakis, Melinos Averkiou, Subhransu Maji,
and Siddhartha Chaudhuri. 3d shape segmentation with pro-
jective convolutional networks. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition,
pages 3779–3788, 2017. 2
[22] Taichi Kin, Hirofumi Nakatomi, Masaaki Shojima, Minoru
Tanaka, Kenji Ino, Harushi Mori, Akira Kunimatsu, Hiroshi
Oyama, and Nobuhito Saito. A new strategic neurosurgical
planning tool for brainstem cavernous malformations using
interactive computer graphics with multimodal fusion im-
ages. Journal of neurosurgery, 117(1):78–88, 2012. 4
[23] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis
Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa,
Denis Zorin, and Daniele Panozzo. Abc: A big cad model
dataset for geometric deep learning. In The IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), June
2019. 2
[24] Fabien Lareyre, Cedric Adam, Marion Carrier, Carine Dom-
merc, Claude Mialhe, and Juliette Raffort. A fully automated
pipeline for mining abdominal aortic aneurysm using image
segmentation. Scientific reports, 9(1):1–14, 2019. 1
[25] Max WK Law and Albert CS Chung. Vessel and intracra-
nial aneurysm segmentation using multi-range filters and lo-
cal variances. In International Conference on Medical Image
Computing and Computer-Assisted Intervention, pages 866–
874. Springer, 2007. 5
[26] Truc Le and Ye Duan. Pointgrid: A deep network for 3d
shape understanding. In Proceedings of the IEEE conference
on computer vision and pattern recognition, pages 9204–
9214, 2018. 2, 5
[27] Jiaxin Li, Ben M Chen, and Gim Hee Lee. So-net: Self-
organizing network for point cloud analysis. In Proceed-
2664
ings of the IEEE conference on computer vision and pattern
recognition, pages 9397–9406, 2018. 3, 5
[28] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di,
and Baoquan Chen. Pointcnn: Convolution on x-transformed
points. In Advances in Neural Information Processing Sys-
tems, pages 820–830, 2018. 3, 5
[29] Jinxian Liu, Bingbing Ni, Caiyuan Li, Jiancheng Yang, and
Qi Tian. Dynamic points agglomeration for hierarchical
point sets learning. In The IEEE International Conference
on Computer Vision (ICCV), October 2019. 3
[30] Karen Lopez-Linares, Inmaculada Garcıa, Ainhoa Garcıa-
Familiar, Ivan Macıa, and Miguel A Gonzalez Ballester. 3d
convolutional neural network for abdominal aortic aneurysm
segmentation. arXiv preprint arXiv:1903.00879, 2019. 1
[31] Haggai Maron, Meirav Galun, Noam Aigerman, Miri Trope,
Nadav Dym, Ersin Yumer, Vladimir G Kim, and Yaron Lip-
man. Convolutional neural networks on surfaces via seam-
less toric covers. ACM Trans. Graph., 36(4):71–1, 2017. 2,
3
[32] Jeong Joon Park, Peter Florence, Julian Straub, Richard
Newcombe, and Steven Lovegrove. Deepsdf: Learning con-
tinuous signed distance functions for shape representation.
In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 165–174, 2019. 5
[33] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas.
Pointnet: Deep learning on point sets for 3d classification
and segmentation. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pages 652–660,
2017. 2, 5
[34] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J
Guibas. Pointnet++: Deep hierarchical feature learning on
point sets in a metric space. In Advances in neural informa-
tion processing systems, pages 5099–5108, 2017. 2, 5
[35] Pranav Rajpurkar, Jeremy Irvin, Aarti Bagul, Daisy Ding,
Tony Duan, Hershel Mehta, Brandon Yang, Kaylie Zhu, Dil-
lon Laird, Robyn L Ball, et al. Mura: Large dataset for ab-
normality detection in musculoskeletal radiographs. arXiv
preprint arXiv:1712.06957, 2017. 2
[36] Safa Salahat, Ahmed Soliman, Tim McGloughlin, Naoufel
Werghi, and Ayman El-Baz. Segmentation of abdominal aor-
tic aneurysm (aaa) based on topology prior model. In Annual
Conference on Medical Image Understanding and Analysis,
pages 219–228. Springer, 2017. 5
[37] T Sichtermann, A Faron, R Sijben, N Teichert, J Freiherr,
and M Wiesmann. Deep learning–based detection of in-
tracranial aneurysms in 3d tof-mra. American Journal of
Neuroradiology, 40(1):25–32, 2019. 1
[38] Amber L. Simpson, Michela Antonelli, Spyridon Bakas,
Michel Bilello, Keyvan Farahani, Bram van Ginneken, An-
nette Kopp-Schneider, Bennett A. Landman, Geert J. S. Lit-
jens, Bjoern H. Menze, Olaf Ronneberger, Ronald M. Sum-
mers, Patrick Bilic, Patrick Ferdinand Christ, Richard K. G.
Do, Marc Gollub, Jennifer Golia-Pernicka, Stephan Heck-
ers, William R. Jarnagin, Maureen McHugo, Sandy Napel,
Eugene Vorontsov, Lena Maier-Hein, and M. Jorge Cardoso.
A large annotated medical image dataset for the develop-
ment and evaluation of segmentation algorithms. CoRR,
abs/1902.09063, 2019. 2
[39] Hang Su, Varun Jampani, Deqing Sun, Subhransu Maji,
Evangelos Kalogerakis, Ming-Hsuan Yang, and Jan Kautz.
Splatnet: Sparse lattice networks for point cloud processing.
In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 2530–2539, 2018. 3
[40] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik
Learned-Miller. Multi-view convolutional neural networks
for 3d shape recognition. In Proceedings of the IEEE in-
ternational conference on computer vision, pages 945–953,
2015. 2
[41] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud,
Beatriz Marcotegui, Francois Goulette, and Leonidas J
Guibas. Kpconv: Flexible and deformable convolution for
point clouds. In Proceedings of the IEEE International Con-
ference on Computer Vision, pages 6411–6420, 2019. 3
[42] B. van Ginneken, M.B. Stegmann, and M. Loog. Segmenta-
tion of anatomical structures in chest radiographs using su-
pervised methods: a comparative study on a public database.
Medical Image Analysis, 10(1):19–40, 2006. 2
[43] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun,
and Xin Tong. O-cnn: Octree-based convolutional neu-
ral networks for 3d shape analysis. ACM Transactions on
Graphics (TOG), 36(4):72, 2017. 2
[44] Yunhai Wang, Shmulik Asafi, Oliver Van Kaick, Hao Zhang,
Daniel Cohen-Or, and Baoquan Chen. Active co-analysis
of a set of shapes. ACM Transactions on Graphics (TOG),
31(6):165, 2012. 2
[45] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma,
Michael M Bronstein, and Justin M Solomon. Dynamic
graph cnn for learning on point clouds. ACM Transactions
on Graphics (TOG), 38(5):146, 2019. 3, 5
[46] Zongji Wang and Feng Lu. Voxsegnet: Volumetric cnns for
semantic part segmentation of 3d shapes. IEEE transactions
on visualization and computer graphics, 2019. 2
[47] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep
convolutional networks on 3d point clouds. In Proceedings
of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 9621–9630, 2019. 3, 5
[48] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Lin-
guang Zhang, Xiaoou Tang, and Jianxiong Xiao. 3d
shapenets: A deep representation for volumetric shapes. In
Proceedings of the IEEE conference on computer vision and
pattern recognition, pages 1912–1920, 2015. 2
[49] Yifan Xu, Tianqi Fan, Mingye Xu, Long Zeng, and Yu Qiao.
Spidercnn: Deep learning on point sets with parameterized
convolutional filters. In Proceedings of the European Con-
ference on Computer Vision (ECCV), pages 87–102, 2018.
3, 5
[50] Jiancheng Yang, Qiang Zhang, Bingbing Ni, Linguo Li,
Jinxian Liu, Mengdie Zhou, and Qi Tian. Modeling point
clouds with self-attention and gumbel subset sampling. In
Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 3323–3332, 2019. 3
[51] Yongheng Zhao, Tolga Birdal, Haowen Deng, and Federico
Tombari. 3d point capsule networks. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recogni-
tion, pages 1009–1018, 2019. 3, 6
2665
[52] Qingnan Zhou and Alec Jacobson. Thingi10k: A
dataset of 10,000 3d-printing models. arXiv preprint
arXiv:1605.04797, 2016. 2
2666